A conclusion based on all six choices together. Use it as your list of requirements when you talk to AI vendors.
In short
Your profile points to an AI platform that runs entirely in your own environment, with rules and documentation built in from day one.
Your profile points to AI that runs entirely in your own environment, on models you operate yourselves.
Your profile points to a European-hosted AI platform with rules and documentation built in, rather than a standard AI subscription.
Your profile points to AI hosted under European control, with open models for anything sensitive.
A mainstream AI service can work for you, but only with a governance layer on top that enforces your rules and documents how AI is used.
A mainstream AI service that keeps your data in the EU fits your needs. Most of the effort will go into connecting it to your systems.
A standard AI assistant covers your needs today. Start simple, and take this check again as use grows.
AI models
Only open models you run on your own servers, such as gpt-oss-120b, Qwen, Mistral or Gemma. Choose the largest your hardware allows, because system work and multi-step tasks need strong models. ChatGPT, Claude and Gemini are ruled out.
Only open models you run on your own servers, such as gpt-oss, Qwen, Mistral or Gemma. Mid-size models are enough for writing, summaries and search. ChatGPT, Claude and Gemini are ruled out.
Large open models hosted in the EU, such as gpt-oss-120b, Qwen or Mistral Large, for everything involving sensitive data. ChatGPT, Claude or Gemini only for tasks without sensitive data, routed through a gateway that enforces this.
Open models hosted in the EU, such as gpt-oss, Qwen, Mistral or Gemma, cover your tasks. Services from outside the EU only for tasks without sensitive data.
The strongest models, such as ChatGPT, Claude or Gemini, on plans that keep your data in the EU. Check each vendor, as EU data residency isn't offered for every product.
Any mainstream assistant on a plan that keeps your data in the EU and doesn't use it for training: Copilot, ChatGPT, Gemini or an EU-hosted open model.
The strongest models, such as ChatGPT, Claude or Gemini, give the best results for system work and agents.
Any mainstream assistant will do: Copilot, ChatGPT, Claude, Gemini, or an open model.
Hosting and data
Your own servers or an isolated data centre, with your own encryption keys. Plan for GPU hardware and people to run it, or a partner who operates it inside your walls.
A European-owned or sovereign cloud, so no company outside the EU can be forced to hand over your data. Ask vendors who owns the infrastructure and which country's law applies.
Data stored and processed in the EU, a data processing agreement, and a written promise that your data is not used for training.
A standard business contract and a data processing agreement are enough.
Rules and documentation
You need a governance platform that enforces your rules automatically and keeps records that can't be changed of every question, answer, AI version and decision. The admin settings built into AI tools are not enough at this level.
You need rules per task and records good enough for an audit: a list of every AI use, automatic checks on what goes in and comes out, and logs for selected tasks.
You need rules per task with automatic checks, and a list of approved AI uses. Your documentation needs are lighter, so standard usage logs will do.
Your rules are simple, but you must be able to show auditors how AI is used. Keep a list of AI uses, and log questions and answers for selected tasks.
One company login, a list of approved tools and regular usage reports cover your needs.
An AI policy, staff training and the AI tool's built-in settings are enough.
Connection to your systems
Plan for an integration programme. AI that works across several systems needs coordination, monitoring, test environments and human sign-off on important steps.
Budget for custom links to your case, finance or customer systems, and make sure AI only sees what each user is allowed to see.
Standard links to your document storage and intranet, where people only see what they already have access to.
No integration needed. People work directly in the AI tool.
Cost
Running AI on your own servers has a high fixed starting cost for a small number of users. Consider a shared or partner-operated setup until use grows.
At your scale, dedicated capacity in Europe or in your own data centre gives a predictable price per task. Compare it with paying per use before you commit.
Reserve or own computing capacity for a predictable price per task, and follow cost per task as a key figure.
Pay per use, with budgets and spending limits per department, and send simple tasks to cheaper models.
Licences for the whole organisation, with cost followed per department. Compare paying per user with paying per use.
A price per user or pay per use keeps the cost low and predictable.
Watch out for
The most capable AI models can't run on your own servers. Expect lower quality on the hardest multi-step tasks, or allow tasks without sensitive data to use external services.
Agents and system work on open models hosted in the EU is possible, but quality varies. Test on your own tasks before you commit.
Rules you can't document are hard to prove. Even simple logging helps when someone asks how AI was used.
AI that completes tasks on its own needs clear rules for what it may do, even if other AI use is relaxed.
AI that updates your systems needs more than common sense. Decide which tasks it may do before you connect it.
Thousands of users with only general guidelines makes unapproved AI use hard to control. Consider approved tools and one company login.
Next step
Share this summary with the vendors you are considering, and ask each one to show how they meet every point. Points they can't meet show where you need a partner or a different solution.